University of Iowa Iowa Research Online Theses and Dissertations Fall 2009 Wind turbine vibration study: a data driven methodology Zijun Zhang University of Iowa Copyright 2009 Zijun Zhang This thesis is available at Iowa Research Online: http://ir.uiowa.edu/etd/454 Recommended Citation Zhang, Zijun. "Wind turbine vibration study: a data driven methodology." MS (Master of Science) thesis, University of Iowa, 2009. http://ir.uiowa.edu/etd/454. Follow this and additional works at: http://ir.uiowa.edu/etd Part of the Industrial Engineering Commons
WIND TURBINE VIBRATION STUDY: A DATA-DRIVEN METHODOLOGY by Zijun Zhang A thesis submitted in partial fulfillment of the requirements for the Master of Science degree in Industrial Engineering in the Graduate College of The University of Iowa December 2009 Thesis Supervisor: Professor Andrew Kusiak
Copyright by ZIJUN ZHANG 2009 All Rights Reserved
Graduate College The University of Iowa Iowa City, Iowa CERTIFICATE OF APPROVAL MASTER S THESIS This is to certify that the Master s Thesis of Zijun Zhang has been approved by the Examining Committee for the thesis requirement for the Master of Science degree in Industrial Engineering at the December 2009 graduation. Thesis Committee: Andrew Kusiak, Thesis Supervisor Yong Chen Kate Cowles
To My Parents and Family ii
Hope is a waking dream. Aristotle iii
ACKNOWLEDGMENTS I would like to express my sincere gratitude to my advisor Professor Andrew Kusiak, for his devotion to this research. He has been the most instrumental person for my academic and research achievements. He provided the motivation, encouragement, guidance and advice which have prepared me for the challenges of future life. I was fortunately exposed to industrial applications while working in the Intelligent Systems Laboratory. This invaluable experience has allowed me to maintain a balance between theory and practice leading to realistic solutions. I would like to thank Professor Yong Chen and Professor Kate Cowles for serving on my Thesis Committee and providing valuable suggestions and feedback on my research. I am also grateful for the financial support from Iowa Energy Center and MidAmerican Energy Company. The energy experts from company and Iowa Energy Center have extended invaluable information for this research. I thank all the members of the Intelligent Systems Laboratory who have worked with me and provided advice, reviews and suggestions. Special thanks to my colleagues: Dr. Zhe Song, who worked with me to solve challenging problems in wind energy domain; Wenyan Li, who shared her research experience with me; Mingyang Li, who enhanced my research capability through frequent communication and collaboration; and Robert A. Hamel, who discussed with me in wind energy topics and provided access to industrial data. And finally, and most importantly, I would like to express my sincere gratitude to my parents, who solidly supported me in my academic pursuit. iv
ABSTRACT Vibrations of a wind turbine have a negative impact on its performance and therefore approaches to effectively control turbines are sought by wind industry. The body of previous research on wind turbine vibrations has focused on physics-based models. Such models come with limitations as some ideal assumptions do not reflect reality. In this Thesis a data-driven approach to analyze the wind turbine vibrations is introduced. Improvements in the data collection of information system allow collection of large volumes of industrial process data. Although the sufficient information is contained in collected data, they cannot be fully utilized to solve the challenging industrial modeling issues. Data-mining is a novel science offers platform to identify models or recognize patterns from large data set. Various successful applications of data mining proved its capability in extracting models accurately describing the processes of interest. The vibrations of a wind turbine originate at various sources. This Thesis focuses on mitigating vibrations with wind turbine control. Data mining algorithms are utilized to construct vibration models of a wind turbine that are represented by two parameters, drive train acceleration and tower acceleration. An evolutionary strategy algorithm is employed to optimize the wind turbine performance expressed with three objectives, power generation, vibration of wind turbine drive train, and vibration of wind turbine tower. The methodology presented in the Thesis is applicable to industrial processes other than wind industry. v
TABLE OF CONTENTS LIST OF TABLES...x LIST OF FIGURES... xiii CHAPTER1. INTRODUCTION...1 1.1 Review of wind turbine vibration research...2 1.2 Review of approaches for building predictive models...3 1.3 Computational intelligence and optimization...4 1.4 Thesis structure...5 CHAPTER 2. ANALYSIS OF WIND TURBINE VIBRATION...7 2.1 Introduction...7 2.2 Data description...7 2.3 Data pre-processing...8 2.4 Data analysis of wind turbine vibration in time domain...11 2.4.1 Analysis of data set with wind speed between 3.5m/s and 7m/s...12 2.4.2 Analysis of data set with wind speed between 7m/s and 12m/s...17 2.4.3 Analysis of data set with wind speed higher than 12m/s...20 2.5 Data analysis of wind turbine vibration in frequency domain...23 2.6 Summary...24 CHAPTER 3. MODELING WIND TURBINE VIBRATIONS BASED ON DATA DRIVEN APPROACH...26 3.1 Introduction...26 3.2 Wind-speed based scenarios...27 3.3 Wavelet analysis...27 3.4 Data-driven models...29 3.5 Case study...32 3.6 Summary...40 CHAPTER 4. OPTIMIZATION OF WIND TURBINE PERFORMANCE WITH DATA DRIVEN MODELS...42 4.1 Introduction...42 4.2 Modeling wind turbine vibrations and power output...44 4.2.1 Data description...44 4.2.2 Data pre-processing...45 4.2.3 Wind turbine vibration model...45 4.2.4 Power output model...46 4.2.5 Validation of the models...47 4.3 Multi-objective optimization model...52 4.4 Solving the multi-objective optimization model...55 4.4.1 Strength Pareto Evolutionary Algorithm...55 4.4.2 Tuning parameters of the evolutionary strategy algorithm...57 vi
4.5 Computational results...62 4.5.1 Single-point optimization...62 4.5.2 Multi-point optimization...64 4.5.3 Analysis of computational results...70 4.6. Summary...71 CHAPTER 5. CONCLUSION...74 APPENDIX A. FIGURES ILLUSTRATING PREDICTION PERFORMANCE OF TURBINE 1...76 APPENDIX B. FIGURES ILLUSTRATING PREDICTION PERFORMANCE OF TURBINE 2...93 REFERENCES...110 vii
LIST OF TABLES Table 2.1 Sample data set...8 Table 2.2 Three data subsets...10 Table 2.3 Table 2.4 Table 2.5 Table 2.6 Table 2.7 Table 2.8 Table 2.9 Ranking produced by predictor importance analysis for two turbines for data partition 1...14 Rankings produced by the global sensitivity analysis for two turbines for data partition 1...15 Rankings produced by the correlation coefficient analysis for two turbines for data partition 1...15 Ranking produced by the predictor importance analysis for two turbines in data partition 2...18 Rankings produced by the global sensitivity analysis for two turbines in data partition 2...19 Rankings produced by the correlation coefficient analysis for two turbines in data partition 2...19 Ranking produced by the predictor importance analysis for two turbines in data partition 3...22 Table 2.10 Ranking produced by the global sensitivity analysis for two turbines in data partition 3...23 Table 2.11 Ranking produced by the correlation coefficient analysis for two turbines in data partition 3...23 Table 3.1 Difference between the mean of the original and the denoised drive train acceleration...28 Table 3.2 Training results of the neural network model...28 Table 3.3 Test results of the neural network model...29 Table 3.4 Performance of five classifiers for predicting drive train acceleration...31 Table 3.5 Performance of five classifiers for predicting tower acceleration...31 Table 3.6 Feature descriptions...32 Table 3.7 Test results for wind turbine vibration produced by the neural network model...34 Table 3.8 Test results for wind turbine vibration of data set of Turbine 2...37 Table 4.1 Sample data set of 10-s data collected from SCADA system....44 viii
Table 4.2 Sample 1-min data computed based on the 10-s data...44 Table 4.3 Test results of the NN models for 10-s data...48 Table 4.4 Testing results of the NN models for 1-min data...50 Table 4.5 Description of parameters...53 Table 4.6 Correlation coefficients between turbine parameters...54 Table 4.7 Two experiments for tuning selection pressure and population size...58 Table 4.8 Convergence for 10 values of the selection pressure in experiment 1...59 Table 4.9 Convergence for 10 values of the selection pressure in experiment 2...60 Table 4.10 Convergence of the ES algorithm for two populations of experiment 1...61 Table 4.11 Convergence of the ES algorithm for two populations of experiment 2...62 Table 4.12 Partial solution set generated by the evolutionary strategy algorithm...63 Table 4.13 Gains in vibration reductions of the drive train for Case 1...65 Table 4.14 Gain in reduction tower vibrations for Case 2...67 Table 4.15 Gains in power output for control strategy of Case 3...68 Table 4.16 Comparison of computational results for 10-s data set and 1-min data set over 10 min horizon...71 ix
LIST OF FIGURES Figure 1.1 The Thesis structure...5 Figure 2.1 Torque histogram for data partition 1 of turbine 1...10 Figure 2.2 Torque histogram for data partition 1 of turbine 1 after sampling...11 Figure 2.3 Histogram of the blade pitch angle rate of turbine 1 in data partition 1...13 Figure 2.4 Histogram of the blade pitch angle rate of turbine 2 in data partition 1...13 Figure 2.5 Histogram of the blade pitch angle rate of turbine 1 in data partition 2...17 Figure 2.6 Histogram of the blade pitch angle rate of turbine 2 in data partition 2...18 Figure 2.7 Histogram of torque rate of turbine 1 in data partition 3...21 Figure 2.8 Histogram of torque rate of turbine 2 in data partition 3...21 Figure 2.9 Figure 3.1 Figure 3.2 Figure 3.3 Figure 3.4 Figure 3.5 Figure 3.6 Figure 3.7 Figure 3.8 Spectrum from 0 to 0.05 Hz of the drive train acceleration of Turbine 1...24 Scatter plot of the observed and predicted values of drive train acceleration for the first 200 points of turbine 1 in scenario 1...35 Scatter plot of the observed and predicted values of tower acceleration for the first 200 points of turbine 1 in scenario 1...35 Run-chart of the observed and predicted values of drive train acceleration for the first 200 points of turbine 1 in scenario 1...36 Run-chart of the observed and predicted values of tower acceleration for the first 200 points of turbine 1 in scenario 1...36 Scatter plot of the observed and predicted values of the drive train acceleration for the first 200 points of turbine 2 in scenario 1...38 Scatter plot of the observed and predicted values of the tower acceleration for the first 200 points of turbine 2 in scenario 1...39 Run-chart of the observed and predicted values of the drive train acceleration for the first 200 points of turbine 2 in scenario 1...39 Run-chart of the observed and predicted values of the tower acceleration for the first 200 points of turbine 2 in scenario 1...40 Figure 4.1 Power curve of a 1.5 MW turbine...47 Figure 4.2 The first 50 test points of the drive train acceleration for 10-s data...49 Figure 4.3 The first 50 test points of the tower acceleration for 10-s data...49 x
Figure 4.4 The first 50 test points of the power output for 10-s data...50 Figure 4.5 The first 50 test points of the drive train accelerations for 1-min data...51 Figure 4.6 The first 50 test points of the tower acceleration for 1-min data...51 Figure 4.7 The first 50 test points of the power output 1-min data...52 Figure 4.8 Solution of the elite set in a 3-dimensional space...64 Figure 4.9 The optimized and original drive train acceleration of Case 1 for 10-s data...65 Figure 4.10 The computed and original torque value of Case 1 for 10-s data...66 Figure 4.11 The computed and original average blade pitch angle of Case 1 for 10-s data...66 Figure 4.12 The optimized and original tower acceleration of Case 2 for 10-s data...67 Figure 4.13 The computed and original torque value of Case 2 for 10-s data....67 Figure 4.14 The computed and original blade pitch angle of Case 2 for 10-s data...68 Figure 4.15 The optimized and original power output of Case 3 for 10-s data...69 Figure 4.16 The computed and original torque value of Case 3 for 10-s data...69 Figure 4.17 The computed and original mean blade pitch angle of Case 3 for 10-s data...70 Figure A.1 Figure A.2 Figure A.3 Figure A.4 Figure A.5 Figure A.6 Figure A.7 Figure A.8 Scatter plot of the observed and predicted values of drive train acceleration for the first 200 points of Turbine 1 in Scenario 2...76 Scatter plot of the observed and predicted values of tower acceleration for the first 200 points of Turbine 1 in Scenario 2...77 Run-chart of the observed and predicted values of drive train acceleration for the first 200 points of Turbine 1 in Scenario 2...77 Run-chart of the observed and predicted values of tower acceleration for the first 200 points of Turbine 1 in Scenario 2...78 Scatter plot of the observed and predicted values of drive train acceleration for the first 200 points of Turbine 1 in Scenario 3...78 Scatter plot of the observed and predicted values of tower acceleration for the first 200 points of Turbine 1 in Scenario 3...79 Run-chart of the observed and predicted values of drive train acceleration for the first 200 points of Turbine 1 in Scenario 3...79 Run-chart of the observed and predicted values of tower acceleration for the first 200 points of Turbine 1 in Scenario 3...80 xi
Figure A.9 Scatter plot of the observed and predicted values of drive train acceleration for the first 200 points of Turbine 1 in Scenario 4...80 Figure A.10 Scatter plot of the observed and predicted values of tower acceleration for the first 200 points of Turbine 1 in Scenario 4...81 Figure A.11 Run-chart of the observed and predicted values of drive train acceleration for the first 200 points of Turbine 1 in Scenario 4...81 Figure A.12 Run-chart of the observed and predicted values of tower acceleration for the first 200 points of Turbine 1 in Scenario 4...82 Figure A.13 Scatter plot of the observed and predicted values of drive train acceleration for the first 200 points of Turbine 1 in Scenario 5...82 Figure A.14 Scatter plot of the observed and predicted values of tower acceleration for the first 200 points of Turbine 1 in Scenario 5...83 Figure A.15 Run-chart of the observed and predicted values of drive train acceleration for the first 200 points of Turbine 1 in Scenario 5...83 Figure A.16 Run-chart of the observed and predicted values of tower acceleration for the first 200 points of Turbine 1 in Scenario 5...84 Figure A.17 Scatter plot of the observed and predicted values of drive train acceleration for the first 200 points of Turbine 1 in Scenario 6...84 Figure A.18 Scatter plot of the observed and predicted values of tower acceleration for the first 200 points of Turbine 1 in Scenario 6...85 Figure A.19 Run-chart of the observed and predicted values of drive train acceleration for the first 200 points of Turbine 1 in Scenario 6...85 Figure A.20 Run-chart of the observed and predicted values of tower acceleration for the first 200 points of Turbine 1 in Scenario 6...86 Figure A.21 Scatter plot of the observed and predicted values of drive train acceleration for the first 200 points of Turbine 1 in Scenario 7...86 Figure A.22 Scatter plot of the observed and predicted values of tower acceleration for the first 200 points of Turbine 1 in Scenario 7...87 Figure A.23 Run-chart of the observed and predicted values of drive train acceleration for the first 200 points of Turbine 1 in Scenario 7...87 Figure A.24 Run-chart of the observed and predicted values of tower acceleration for the first 200 points of Turbine 1 in Scenario 7...88 Figure A.25 Scatter plot of the observed and predicted values of drive train acceleration for the first 200 points of Turbine 1 in Scenario 8...88 Figure A.26 Scatter plot of the observed and predicted values of tower acceleration for the first 200 points of Turbine 1 in Scenario 8...89 xii
Figure A.27 Run-chart of the observed and predicted values of drive train acceleration for the first 200 points of Turbine 1 in Scenario 8...89 Figure A.28 Run-chart of the observed and predicted values of tower acceleration for the first 200 points of Turbine 1 in Scenario 8...90 Figure A.29 Scatter plot of the observed and predicted values of drive train acceleration for the first 200 points of Turbine 1 in Scenario 9...90 Figure A.30 Scatter plot of the observed and predicted values of tower acceleration for the first 200 points of Turbine 1 in Scenario 9...91 Figure A.31 Run-chart of the observed and predicted values of drive train acceleration for the first 200 points of Turbine 1 in Scenario 9...91 Figure A.32 Run-chart of the observed and predicted values of tower acceleration for the first 200 points of Turbine 1 in Scenario 9...92 Figure B.1 Figure B.2 Figure B.3 Figure B.4 Figure B.5 Figure B.6 Figure B.7 Figure B.8 Figure B.9 Scatter plot of the observed and predicted values of drive train acceleration for the first 200 points of Turbine 2 in Scenario 2...93 Scatter plot of the observed and predicted values of tower acceleration for the first 200 points of Turbine 2 in Scenario 2...94 Run-chart of the observed and predicted values of drive train acceleration for the first 200 points of Turbine 2 in Scenario 2...94 Run-chart of the observed and predicted values of tower acceleration for the first 200 points of Turbine 2 in Scenario 2...95 Scatter plot of the observed and predicted values of drive train acceleration for the first 200 points of Turbine 2 in Scenario 3...95 Scatter plot of the observed and predicted values of tower acceleration for the first 200 points of Turbine 2 in Scenario 3...96 Run-chart of the observed and predicted values of drive train acceleration for the first 200 points of Turbine 2 in Scenario 3...96 Run-chart of the observed and predicted values of tower acceleration for the first 200 points of Turbine 2 in Scenario 3...97 Scatter plot of the observed and predicted values of drive train acceleration for the first 200 points of Turbine 2 in Scenario 4...97 Figure B.10 Scatter plot of the observed and predicted values of tower acceleration for the first 200 points of Turbine 2 in Scenario 4...98 Figure B.11 Run-chart of the observed and predicted values of drive train acceleration for the first 200 points of Turbine 2 in Scenario 4...98 Figure B.12 Run-chart of the observed and predicted values of tower acceleration for the first 200 points of Turbine 2 in Scenario 4...99 xiii
Figure B.13 Scatter plot of the observed and predicted values of drive train acceleration for the first 200 points of Turbine 2 in Scenario 5...99 Figure B.14 Scatter plot of the observed and predicted values of tower acceleration for the first 200 points of Turbine 2 in Scenario 5...100 Figure B.15 Run-chart of the observed and predicted values of drive train acceleration for the first 200 points of Turbine 2 in Scenario 5...100 Figure B.16 Run-chart of the observed and predicted values of tower acceleration for the first 200 points of Turbine 2 in Scenario 5...101 Figure B.17 Scatter plot of the observed and predicted values of drive train acceleration for the first 200 points of Turbine 2 in Scenario 6...101 Figure B.18 Scatter plot of the observed and predicted values of tower acceleration for the first 200 points of Turbine 2 in Scenario 6...102 Figure B.19 Run-chart of the observed and predicted values of drive train acceleration for the first 200 points of Turbine 2 in Scenario 6...102 Figure B.20 Run-chart of the observed and predicted values of tower acceleration for the first 200 points of Turbine 2 in Scenario 6...103 Figure B.21 Scatter plot of the observed and predicted values of drive train acceleration for the first 200 points of Turbine 2 in Scenario 7...103 Figure B.22 Scatter plot of the observed and predicted values of tower acceleration for the first 200 points of Turbine 2 in Scenario 7...104 Figure B.23 Run-chart of the observed and predicted values of drive train acceleration for the first 200 points of Turbine 2 in Scenario 7...104 Figure B.24 Run-chart of the observed and predicted values of tower acceleration for the first 200 points of Turbine 2 in Scenario 7...105 Figure B.25 Scatter plot of the observed and predicted values of drive train acceleration for the first 200 points of Turbine 2 in Scenario 8...105 Figure B.26 Scatter plot of the observed and predicted values of tower acceleration for the first 200 points of Turbine 2 in Scenario 8...106 Figure B.27 Run-chart of the observed and predicted values of drive train acceleration for the first 200 points of Turbine 2 in Scenario 8...106 Figure B.28 Run-chart of the observed and predicted values of tower acceleration for the first 200 points of Turbine 2 in Scenario 8...107 Figure B.29 Scatter plot of the observed and predicted values of drive train acceleration for the first 200 points of Turbine 2 in Scenario 9...107 Figure B.30 Scatter plot of the observed and predicted values of tower acceleration for the first 200 points of Turbine 2 in Scenario 9...108 xiv
Figure B.31 Run-chart of the observed and predicted values of drive train acceleration for the first 200 points of Turbine 2 in Scenario 9...108 Figure B.32 Run-chart of the observed and predicted values of tower acceleration for the first 200 points of Turbine 2 in Scenario 9...109 xv
1 CHAPTER 1. INTRODUCTION Recent years have seen growing interest in renewable energy. This increase is driven, in part, by growing awareness of the energy cost, climate changes, supply uncertainty, and environment concerns. Wind energy is recognized as one of the most important sources of renewable energy and this awareness has translated in expansion of investments in this area. In 2008, the U.S. Department of Energy has produced a report aiming at increasing contribution of wind energy to the electricity supply to 20% by 2030. However, challenging issues, such as higher operation, maintenance and market costs than other conventional energy sources in many areas across the country, create a great barrier on this road. To accomplish this ambitious goal, numerous questions of wind energy need to be addressed, including providing control strategies to mitigate wind turbine vibrations. Research in conventional power systems has introduced numerous simulation models for analysis of different operational scenarios of those systems. Yet, a commercial wind farm includes a large number (dozens to hundreds) of megawatt-class turbines with static and dynamic characteristics that differ from the conventional power plants. Therefore, the modeling template developed for conventional power generating facilities is not compatible with modeling wind turbines. Novel methodologies to model wind turbine systems are needed by wind energy industry. The research in wind energy has intensified in recent years. Areas with the most research progress include the design of wind turbines [1, 2], the design and reliability of wind farms [3, 4, 5], the control of wind turbines [6, 7, 8, 9], wind energy conversion [10, 11], the prediction of wind power [12, 13], and condition monitoring of wind turbines [14, 15]. Although numerous novel modeling methods were addressed in the literature, the major focus was on developing accurate power prediction models compare to other topics. Wind turbine vibrations impact performance and life-cycle of wind turbines and
2 therefore it deserves further studies. Mitigating the vibrations of a wind turbine can potentially prevent material fatigue, reduce number of component failures, improve power quality and extend life-cycle of some components, such as gearbox. This in turn translates into increased turbine availability and reduced maintenance costs. The goal of the Thesis is to analyze wind turbine vibrations and developing nonlinear and nonparametric models to optimize wind turbine performance in considering three objectives, maximization of the power produced by a wind turbine, and reduction of vibrations of the turbine s drive train and tower. Researchers introduced models to predict wind power and describe vibrations of components of a wind turbine. Nevertheless, the majority of the published research falls into parametric and physics-based models. It is widely recognized such models usually involve assumptions, and therefore they may not adequately represent reality. 1.1 Review of wind turbine vibration research The sources of wind turbine vibrations [16] are diverse and due to the large size of wind turbines, conducting laboratory experiments with such systems is difficult. Thus, the past wind turbine vibration research has primarily focused on the building models based on first principles and simulation. Leithead et al. [17] studied dynamics of variable speed wind turbines and design of models to control wind turbines. Fadaeinedjad et al. [18] investigated the impact of voltage sag on vibration of the wind turbine tower. They used three simulation programs, TrubSim, FAST and Simulink, to model wind turbines. Murtagh et al. [19] investigated control wind turbine vibration by incorporating a passive control device. A passive control method using a tuned mass damper to mitigate vibrations of the blades and tower of a wind turbine was introduced. The research reported in [20] discussed the estimation of aeroelastic damping of operational wind turbine modes based on experiments.
3 The physics-based or simulation based wind turbine vibration models introduced in the past research provided a solid foundation of understanding the nature of wind turbine vibrations. Unfortunately, these models involve assumptions that the industry finds restrictive. Therefore, new and more effective approaches are needed. 1.2 Review of approaches for building predictive models Accurate prediction of wind power is essential for integration of commercial wind farms with the electric grid. Prediction of wind speed is one of the elements of power prediction. Numerous approaches to predicting wind speed have been developed. Louka et al. [21] applied Kalman filters to enhance wind speed prediction using hourly data. Flores et al. [22] employed neural networks for hourly prediction of wind speed and designed a control system for active power generation. El-Fouly et al. [23] developed a linear time-series strategy to predict hourly wind speed and direction. Power prediction researches also have been addressed in various published literature. Damousis et al. [24] used a fuzzy logic model trained by a genetic algorithm to predict wind speed and power over 0.5 to 2 hour horizons. Contaxis et al. [25] introduced an ARMA model to predict power produced by wind turbine in a study of short term scheduling issue. Anahua et al. [26] presented a stochastic differential equation to estimate power by describe power generation as a Markov process. Landberg et al. [27] predict the power produced by a wind farm through establishing a model by using the data from the weather prediction model (HIRLAM) and the local weather model (WASP). In the published literature on wind and power prediction, statistical model, physics-based model and climate model have been widely discussed. However, the published research has mainly concentrated on long-time predictions, e.g., hourly predictions. To control of a wind turbine, accurate short-term power prediction models, such as 1-min or 10-s power prediction models are needed. Due to the higher sampling frequency of data, the stochastic nature of the wind speed was more significant and it
4 makes modeling its behavior a challenge. The past states of wind speed, temperature, humidity, latitude, terrain topography, air pressure and other factors all impact wind speed. Thus, the quality of power predicted for a wind turbine is dependent on the accuracy of wind speed prediction. The previous research [13] has demonstrated suitability of the data-driven approach 1.3 Computational intelligence and optimization New theories and advances in computational intelligence, fuzzy logic, and image processing offer alternatives to model and solve problems in energy systems. Chu et al. [28] applied a neural network to predict the performance index and non-analytical constraints, thus speeding up the trial-and-error approach of finding the optimal operating points optimizing a boiler s combustion process. Rusinowski et al. [29] focused on finding an optimal travelling rate of the grid and an optimal height of the fuel layer. Büche et al. [30] applied an evolutionary computation algorithm to find an optimal design of a burner to reduce NOx emissions as well as pressure fluctuation. Wang et al. [31] applied a naïve intelligent control algorithm to determine the best air supply for a boiler. Cass et al. [32] combined the neural network and evolutionary computation techniques to determine an optimal fuel/air ratio. Kusiak et al. [33] addressed anticipatory control of wind turbine by integrating data mining and evolutionary strategy algorithms. Li et al. [34] introduced a hybrid genetic and immune algorithm to solve the optimization problem of internal electric connection system of large offshore wind farms. Prats et al. [35] applied fuzzy control techniques to improve wind energy capture for variable speed and variable pitch wind turbines. Sareni et al. [36] developed multi-objective genetic algorithm to study the optimal design of a small passive wind turbine generator by considering the size, power generation and other issues.
5 1.4 Thesis structure Figure 1.1 illustrates the structure of the Thesis. Five data mining algorithms, Neural Network, Support Vector Machine, Standard C&R Tree, Boosting Tree and Random Forests are used to construct models for wind turbine vibrations. Chapter 2 presents analysis of wind turbine vibration data collected from SCADA systems in time and frequency domain. In Chapter 3, modeling wind turbine vibrations in drive train and tower by data-mining algorithms are discussed. Chapter 4 presents optimization model of wind turbine performance. Figure 1.1 The Thesis structure
6 In Chapter 2, three approaches, namely the predictor importance analysis, the global sensitivity analysis, and the correlation coefficient analysis, are applied to determine turbine parameters that could potentially mitigate turbine vibrations. In the frequency domain analysis, Fourier analysis transforms time domain data into frequency domain to demonstrate another approach to vibration studies. In Chapter 3, application of wavelet analysis in smoothing data is discussed as sensors to measure drive train acceleration and tower acceleration are noise sensitive. In Chapter 4 a framework for wind turbine performance optimization is presented. Data mining and evolutionary algorithms are integrated to model and solve multiobjective optimization model of maximizing power generation, and reducing vibrations of the drive train and the tower. Tuning parameters of evolutionary strategy algorithm is discussed to improve computational efficiency. Three weight assignment cases are discussed and computational gains in power maximization and vibration reduction are presented.
7 CAPTER 2. ANALYSIS OF WIND TURBINE VIBRATION 2.1 Introduction In this chapter, analysis of wind turbine vibration in the time domain and the frequency domain is introduced. Two parameters the drive train acceleration and the tower acceleration are utilized to represent the wind turbine vibration. The sources of wind turbine vibrations [16] are diverse. The focus of this research is on vibrations attributed to the control of wind turbines, e.g., control of the generator torque and blade pitch angle. Data partitioning strategy is applied to divide data set according to fixed range of wind speed so that impact of wind speed in wind turbine vibration can be mitigated and vibrations associated to control parameters of wind turbine can be emphasized. The basis of the time domain analysis is statistical and data-driven methodologies. Three approaches, namely the predictor importance analysis, the global sensitivity analysis, and the correlation coefficient analysis, are applied to determine turbine parameters that could potentially mitigate turbine vibrations. In the frequency domain analysis, data is transformed from time domain to frequency domain by Fourier analysis and this approach offers an alternative angle to understand wind turbine vibration. 2.2. Data description In this research, data sets collected by the SCADA system at two variable speed 1.5MW turbines of a large wind farm are used. Each data set contains average values of more than 120 parameters, including vibration parameters, all stored at 10-second (10-s) intervals and thus the sampling frequency is 0.1Hz. Although the SCADA system contains values of many parameters, only some of them are of interest to vibration analysis. The literature and domain expertise was used to select a list of parameters that
8 could be potentially relevant to the research discussed in this paper. Table 2.1 illustrates the format of the data used in this research. Table 2.1 Sample data set Observati on No. Time Torque Value [%] Wind Speed [m/s] Drive Train Acceleration [mm/s2] Tower Acceleration [mm/s2] 10/1/08 1 12:00 AM 22.10 5.77 25.67 29.31 10/1/08 2 12:00 AM 22.60 6.45 24.78 30.26 10/1/08 3 12:00 AM 23.10 6.07 23.89 31.21 10/8/08 60482 12:00 AM 0.00 2.74 18.01 29.34 As illustrated in Table 2.1, the values of all parameters contained in the data set, such as torque, wind speed, wind deviation, drive train acceleration, and tower acceleration, are time stamped. The wind turbine vibration is indicated by two important parameters, the drive train acceleration reflecting vibrations of the drive train, and the tower acceleration reflecting vibrations of the tower. The accelerometer measuring the drive train acceleration is attached at the rear bottom of nacelle and the tower acceleration accelerometer is located near the nacelle and tower connection. 2.3 Data pre-processing Since the data set is stored at 10-s intervals, the month-long data set considered in this research is large, and it contains errors caused by malfunction of sensors, mechanical systems, and the data collection system. Those errors usually appear as missing values,
9 values that are out of range, and invalid values. For example, the net power produced by a wind turbine should be a positive number, which is usually between 0 and its rated power. Thus, filtering erroneous values is a significant step in data-driven research. However, once the error logic is discovered, the data cleaning process can be automated. After filtering the errors and invalid values, three derived parameters are created based on the original SCADA data. The first one is the wind deviation (yaw error), which is defined as the difference between the wind direction and the nacelle position. The next two are the rate of change of torque and the rate of change of the pitch angle. The rate of change of torque (referred to as torque rate) is the difference between the current torque value and the torque value at the preceding time 10-s interval (see Eq. (2.1)). The rate of change of pitch angle (referred to as blade pitch angle rate) is the difference between the current pitch angle and the pitch angle preceding the 10-s time interval (see Eq. (2.2)). Torque Rate = Torque Value( t) Torque Value( t 1) (2.1) Pitch Angle Rate = Pitch Angle( t) Pitch Angle( t 1) (2.2) The two derived parameters (see Eqs (2.1) and (2.2)) provide additional information about wind turbine vibrations from the rate of change perspective. In the time domain analysis, the entire data set is used for training models, and it is decomposed into three partitions based on the wind speed values: wind speed in the interval [3.5 m/s, 7 m/s), [7 m/s, 12 m/s), and [>=12 m/s]. This rather arbitrary partitioning is based on the sigmoid shape of power curve and provides a way to isolate the turbine vibrations attributed to both the drive train and the tower from the impact of other factors such as the wind itself, malfunctions of mechanical systems (e.g., shaft misalignments), and so on (see Table 2.2).
10 Table 2.2 Three data subsets Wind Turbine 1 Data Partition No. Wind Speed Number of Data Points 1 [3.5m/s, 7m/s) 77593 10-s observations 2 [7m/s, 12m/s) 103148 10-s observations 3 [>=12m/s] 21525 10-s observations Wind Turbine 2 Data Partition No. Wind Speed Number of Data Points 1 [3.5m/s, 7m/s) 63554 10-s observations 2 [7m/s, 12m/s) 103115 10-s observations 3 [>= 12m/s] 11855 10-s observations Although the volume of data collected at the wind farm is large, some data samples are biased, i.e., some observations included in the population dominate other data points. A typical biased data sample of torque values included in Data Partition 1 of Turbine 1 (see Table 2.2) is illustrated in the histogram of Fig. 2.1. 40000 35000 30000 25000 No of obs 20000 15000 10000 5000 0-8.5-7.1-5.6-4.2-2.8-1.3 0.1 1.6 3.0 4.4 5.9 7.3 8.8 Torque Rate Figure 2.1 Torque histogram for data partition 1 of turbine 1
11 It is obvious from the histogram in Fig. 1 that the torque rates in the interval [- 0.84, 0.84] have a much higher frequency than the values in other intervals. Thus, the number of observations in this interval needs to be reduced from about 36000 to 7000. This has been accomplished with a random sampling without a replacement scheme. The histogram of torque values after sampling is presented in Fig. 2.2. 8000 7000 6000 5000 No of obs 4000 3000 2000 1000 0-10.9-9.0-7.1-5.2-3.2-1.3 0.6 2.5 4.4 6.4 8.3 10.2 Torque Rate Figure 2.2 Torque histogram for data partition 1 of turbine 1 after sampling 2.4 Data analysis of wind turbine vibration in time domain In this research, several parameters measured by sensors or derived from data, such as torque, torque rate, wind speed, wind deviation, blade pitch angle (average of the three measured pitch angles, one for each blade), and the blade pitch rate, are considered as the major factors potentially impacting the turbine vibrations. These parameters are selected mainly based on domain knowledge and study of the wind energy literature [37, 38]. The tower and drive train accelerations are recorded by the SCADA system. As there are two similar measured values offered by the sensor installed on the drive train,
12 the average value of drive train acceleration is considered for simplicity of analysis. Three different data approaches are applied to quantitatively analyze the impact of each of the selected parameters on the turbine vibrations reflected by the drive train acceleration and the tower acceleration. The data analysis approaches include the predictor importance analysis, the global sensitivity analysis, and the correlation coefficient analysis, and they are applied to each of the three data partitions of Table 2.2. Predictor importance is determined by the boosting regression tree algorithm [39, 40]. The predictor importance statistics, e.g., the sum of the squares errors, are computed for each split during the process of building trees, and the best predictor parameter is then is selected. An average statistic is computed over all trees and all splits. The predictor parameter with the highest value is assigned the value of 100, and other parameters are assigned lower values. The global sensitivity analysis ranks the importance of inputs on the model extracted by a neural network approach [41, 42, 43]. It examines the contribution of uncertainty of all inputs to the output of the model simultaneously, rather than individually, to determine the order of parameter importance. The correlation coefficient [44] is a statistical approach to analyze the relationship between predictors and the target based on their affinity. 2.4.1 Analysis of data set with wind speed between 3.5m/s and 7m/s For Data Partition 1 of Table 2.2, the wind speed of both turbines is in the interval [3.5 m/s, 7 m/s). Due to the fact that the wind speed is rather low, its impact on the drive train and the tower is likely to be minimal.
13 50000 45000 40000 35000 Number of Observations 30000 25000 20000 15000 10000 5000 0-2.0-1.7-1.4-1.1-0.8-0.5-0.2 0.1 0.4 0.7 1.0 1.3 1.6 1.9 Blade Pitch Angle Rate Figure 2.3 Histogram of the blade pitch angle rate of turbine 1 in data partition 1 35000 30000 25000 Number of Observations 20000 15000 10000 5000 0-2.0-1.7-1.4-1.1-0.8-0.5-0.3 0.0 0.3 0.6 0.9 1.2 1.4 1.7 Blade Pitch Angle Rate Figure 2.4 Histogram of the blade pitch angle rate of turbine 2 in data partition 1
14 It is also known that for low wind speeds, the blade pitch angle remains mostly constant for most pitch controlled turbines as shown in Figure 2.3 and Figure 2.4; therefore, this parameter could be excluded in the analysis. Table 2.3 shows the impact of predictors (measured with the predictor importance) on the drive train acceleration and the average tower acceleration of the two turbines. Table 2.3 Ranking produced by predictor importance analysis for two turbines for data partition 1 Drive Train Acceleration Tower Acceleration Predictor Turbine 1 Predictor Rank Turbine 2 Predictor Rank Turbine 1 Predictor Rank Turbine 2 Predictor Rank Torque Value 100 100 90 95 Torque Rate 79 94 100 98 Wind Deviation 54 57 91 93 Blade Pitch Angle 54 66 67 100 Wind Speed 47 46 51 82 The values of the predictor rank in Table 2.3 are generated by the boosting tree regression algorithm. The predictor importance varies by the predictor (e.g., Torque Value, Torque Rate) and the target (i.e., Drive Train Acceleration, Tower Acceleration). The global sensitivity rankings produced by a neural network are provided in Table 2.4. Although the scale used to rank the predictors is different than the one used in Table 2.3, a higher ranking value indicates that the contribution of the corresponding parameter for making predictions is higher.
15 Table 2.4 Rankings produced by the global sensitivity analysis for two turbines for data partition 1 Drive Train Acceleration Tower Acceleration Predictor Turbine 1 Predictor Rank Turbine 2 Predictor Rank Turbine 1 Predictor Rank Turbine 2 Predictor Rank Torque Value 3.27 3.87 3.27 1.25 Torque Rate 1.55 2.02 1.11 1.04 Wind Deviation 1.03 1.00 1.02 1.01 Blade Pitch Angle 1.02 1.01 1.07 1.00 Wind Speed 1.96 1.33 1.64 1.05 Table 2.5 illustrates the correlation coefficient between predictors and two accelerations. A higher value of the correlation coefficient indicates a stronger dependence between a predictor and the vibration. Table 2.5 Rankings produced by the correlation coefficient analysis for two turbines for data partition 1 Predictor Drive Train Acceleration Turbine 1 Turbine 2 Correlation Correlation Coefficient Coefficient Tower Acceleration Turbine 1 Correlation Coefficient Turbine 2 Correlation Coefficient Torque Value 0.74 0.50 0.30 0.11 Torque Rate -0.23-0.22-0.05 0.01 Wind Deviation 0.14 0.04 0.09 0.05 Blade Pitch Angle -0.36-0.19-0.17-0.06 Wind Speed 0.55 0.28 0.17-0.03 In the boosting tree regression analysis, a higher predictor rank points to a stronger impact of the predictor on a target variable (here vibration). The nature of the global sensitivity analysis is similar to the regression boosting tree analysis. However, the correlation coefficient analysis offers a different concept. A positive correlation coefficient implies that the two variables are positively and linearly correlated, while a
16 negative correlation coefficient indicates the inverse relationship. A higher value of the correlation coefficient indicates a more obvious linear relationship between the corresponding variables. In Table 2.3, the rank of the torque in respect to the drive train acceleration is 100, the highest of all other parameters. In Table 2.4, the rank value of torque is also the highest for both two turbines. In Table 2.5, the correlation coefficient between the torque value and the drive train acceleration is the highest, which means that the vibrations of the drive train are strongly associated with the torque value. These observations indicate that in the speed interval [3.5 m/s, 7 m/s) large values of the torque potentially contribute to higher acceleration of the drive train. The torque rate of change is another variable with a strong impact on vibrations of the drive train of a wind turbine. In the boosting tree regression analysis, the torque rate of change ranked after the torque value for both turbines. In the global sensitivity analysis, it is ranked third for Turbine 1 and second for Turbine 2. The wind speed turns out to be more important for Turbine 1 than for Turbine 2. The correlation coefficient in Table 2.5 provides a different result for the torque rate, as it emphasizes the linear relationship rather than the non-linear relationship between the corresponding variables. In this case, the results of the first two analyses provide more valuable information and indicate that the torque rate of change is another factor (after torque value) strongly associated with the vibrations of the turbine drive train. For the tower vibration, no single parameter consistently scores the highest rank in all three analyses. However, the rank values in Table 2.3 and Table 2.4 indicate that the torque value is more important than most other variables for both turbines. In Table 2.3, the rank of torque value is 90 for Turbine 1 and 95 for Turbine 2. In Table 2.4, the rank of torque value is 3.27 for Turbine 1 and 1.25 for Turbine 2. In conclusion, although the rankings for a turbine tower provided by the three analyses are somehow different, it is apparent that the torque is associated with the vibrations at the turbine tower.
17 2.4.2 Analysis of data set with wind speed between 7m/s and 12m/s In Data Partition 2 of Table 2.1, the wind speed falls in the interval [7 m/s, 12 m/s). Figure 2.5 and 2.6 illustrate the blade pitch angle rate (i.e., the change of pitch angle in the consecutive time points (see Eq. (2)) for the data sets of two turbines. 70000 60000 50000 Number of Observations 40000 30000 20000 10000 0-2.9-2.2-1.6-0.9-0.3 0.4 1.0 1.7 2.3 Blade Pitch Angle Rate Figure 2.5 Histogram of the blade pitch angle rate of turbine 1 in data partition 2
18 60000 50000 Number of Observations 40000 30000 20000 10000 0-2.9-2.3-1.8-1.2-0.7-0.1 0.4 1.0 1.5 2.1 2.7 Figure 2.6 Histogram of the blade pitch angle rate of turbine 2 in data partition 2 The blade pitch angle of two wind turbines (see Figsure 5 and 6) does not significantly change. Table 2.6 shows the results of the predictor importance analysis of two turbines in Data Partition 2. Table 2.7 illustrates the results of the global sensitivity analysis for the two turbines. Table 8 presents the results of the correlation coefficient analysis in this scenario. Table 2.6 Ranking produced by the predictor importance analysis for two turbines in data partition 2 Drive Train Acceleration Tower Acceleration Predictor Importance Analysis Turbine 1 Turbine 2 Turbine 1 Turbine 2 Predictor Rank Predictor Rank Predictor Rank Predictor Rank Torque Value 100 100 100 98 Torque Rate 86 95 84 100 Wind Deviation 42 47 39 48 Blade Pitch Angle 71 76 71 81 Wind Speed 69 55 80 76
19 Table 2.7 Rankings produced by the global sensitivity analysis for two turbines in data partition 2 Drive Train Acceleration Tower Acceleration Global Sensitivity Analysis Turbine 1 Turbine 2 Turbine 1 Turbine 2 Predictor Rank Predictor Rank Predictor Rank Predictor Rank Torque Value 3.87 3.77 1.25 1.17 Torque Rate 2.02 1.90 1.04 1.11 Wind Deviation 1.00 1.00 1.01 1.01 Blade Pitch Angle 1.01 1.02 1.00 1.01 Wind Speed 1.33 1.42 1.05 1.17 Table 2.8 Rankings produced by the correlation coefficient analysis for two turbines in data partition 2 Correlation Coefficient Analysis Drive Train Acceleration Turbine 1 Turbine 2 Correlation Correlation Coefficient Coefficient Tower Acceleration Turbine 1 Correlation Coefficient Turbine 2 Correlation Coefficient Torque Value 0.69 0.51 0.52 0.38 Torque Rate 0.08 0.12 0.08 0.14 Wind Deviation 0.02 0.03 0.00 0.01 Blade Pitch Angle 0.25 0.23 0.23 0.23 Wind Speed 0.63 0.48 0.48 0.35 Torque value is considered as the most important variable in vibration analysis of the drive train. In Table 2.6, the rank of torque value is 100 for both turbines. Table 2.7 confirms the results of Table 2.6. In Table 2.8, the correlation coefficient between the torque value and the drive train acceleration is the highest. These results confirm that the torque value is the most significant parameter related to vibrations of the drive train. Torque rate could be considered as the second most important parameter associated with the drive train vibration. The blade pitch angle could be another parameter potentially
20 causing the wind turbine vibrations, as confirmed by the predictor importance analysis and correlation coefficient analysis. In analyzing tower accelerations, the torque value ranks highest for Turbine 1 (Table 2.6). It also scores the second highest rank (98) for Turbine 2. The global sensitivity analysis (Table 2.7) shows that the torque value is also important, as it gets ranked close to other parameters. In Table 2.8, the correlation coefficient between the torque value and the tower acceleration is the highest. The torque rate and blade pitch angle are also important factors related to the tower acceleration. In the predictor importance analysis, the rank values of the torque rate and the blade pitch angle for both turbines are higher than 70. In Table 2.7, the rank values of the variables besides torque value are similar. In Table 2.8, the blade pitch angle shows a higher correlation with the tower acceleration than the torque rate. In conclusion, although the results from different analyses point to different importance of parameters, the results imply that the torque rate and blade pitch angle are strongly associated with the tower acceleration. 2.4.3 Analysis of data set with wind speed higher than 12m/s In this scenario, all the wind speeds are higher than 12 m/s. As the torque value does not frequently change (see Figure 2.7 and 2.8) it is not considered in the analysis discussed in this section. The predictor importance is reported in Table 2.9; Table 2.10 shows the results of the global sensitivity analysis, and Table 2.11 presents the results of the correlation coefficient analysis. This scenario (speed above 12 m/s) is considered to be high wind speed, and it is likely that some vibrations of the wind turbine are contributed by the wind.